European Journal of Operational Research 156 (2004) 752–768

www.elsevier.com/locate/dsw

O.R. Applications

A crane scheduling method for port container terminals

1
Kap Hwan Kim *, Young-Man Park
Department of Industrial Engineering, Pusan National University, Changjeon-dong, Kumjeong-ku, Pusan 609-735, South Korea
Received 22 January 2001; accepted 21 January 2003

Abstract

This paper discusses the problem of scheduling quay cranes (QCs), the most important equipment in port terminals.
A mixed-integer programming model, which considers various constraints related to the operation of QCs, was for-
mulated. This study proposes a branch and bound (B & B) method to obtain the optimal solution of the QC scheduling
problem and a heuristic search algorithm, called greedy randomized adaptive search procedure (GRASP), to overcome
the computational difficulty of the B & B method. The performance of GRASP is compared with that of the B & B
method.
Ó 2003 Published by Elsevier B.V.

Keywords: Scheduling; Branch and bound; Transportation; Container terminal; Combinatorial optimization

1. Introduction

The productivity of a container terminal can be measured in terms of the productivity of two types of
operations. One type is ship operations, in which containers are discharged from and loaded onto a ship.
The other is receiving and delivery operations, in which containers are transferred to and from outside
trucks. The planning process of ship operations consists of berth planning, quay crane (QC) scheduling (in
practice, called work scheduling), and discharge and load sequencing. During the process of berth planning,
the berthing time and the berthing position of a containership on a wharf are determined. A QC schedule
specifies the service sequence of bays in a ship by each QC and the time schedule for the services. Input data
for QC scheduling consists of a stowage plan of a ship, the ready time of each QC, and a yard map that
shows the storage locations of containers bound for the ship. Finally, during discharge and load se-
quencing, the discharge and load sequence of individual containers are determined based on a QC schedule.
This paper addresses the QC scheduling problem, which is pertinent to the second stage of ship operation
planning.

*
Corresponding author. Tel.: +82-51-510-2419; fax: +82-51-512-7603.
E-mail addresses: kapkim@pusan.ac.kr (K.H. Kim), ymanpark@pusan.ac.kr (Y.-M. Park).
1
Tel.: +82-51-510-1483; fax: +82-51-512-7603.

0377-2217/$ - see front matter Ó 2003 Published by Elsevier B.V.

doi:10.1016/S0377-2217(03)00133-4
 K.H. Kim, Y.-M. Park / European Journal of Operational Research 156 (2004) 752–768 753

Several studies have been conducted to improve the efficiency of ship operations in port container
terminals. The subject of these studies includes the following: berth planning (Brown et al., 1995; Lim,
1998), in which berthing positions and berthing times of vessel are determined; load sequencing (Cho,
1982; Cojeen and Dyke, 1976; Gifford, 1981), in which the loading sequence of individual containers are
determined; and transtainer routing (Kim and Kim, 1999; Narasimhan and Palekar, 2002), in which the
travel route of each transtainer and the number of outbound containers to pick up at each yard-bay are
determined. In addition, Cheung et al. (2002) and Zhang et al. (2002) proposed methods for deploying
yard cranes, methods in which the times and routes of yard crane movements among blocks are deter-
mined.
Daganzo (1989) was the first who discussed the QC scheduling problem. He suggested an algorithm for
determining the number of cranes to assign to ship-bays of multiple vessels. Peterkofsky and Daganzo
(1990) also provided an algorithm for determining the departure times of multiple vessels and the number
of cranes to assign to individual holds of vessels at a specific time segment. They also attempted to minimize
the delay costs. The studies by Daganzo (1989) and Peterkofsky and Daganzo (1990) assumed one task per
ship-bay, a task which needs crane operations during a specific length of time, and did not consider the
interference among QCs or precedence relationships among tasks. In contrast, this study assumes that there
may be multiple tasks involved in a ship-bay, and thus this study divides a task into smaller sizes, compared
to Daganzo (1989) and Peterkofsky and Daganzo (1990).
This study further assumes that the berthing and departure times of a vessel and the operation starting
times of QCs assigned to that vessel are given. This study attempts to determine the schedule of each QC
assigned to a vessel, with the goal of completing all of the ship operations of a vessel as rapidly as possible.
The next section introduces the ship operation and defines the QC scheduling problem. Section 3 pro-
vides a mathematical formulation for the QC scheduling problem. Section 4 proposes a branch and bound
(B & B) algorithm for solving the mathematical formulation. Section 5 applies the heuristic algorithm
GRASP to the QC scheduling problem. Section 6 discusses the results of a numerical experiment, and
Section 7 provides a conclusion.

2. Problem description

The goal of studying the QC scheduling problem is to determine the sequence of discharging and loading
operations that a QC will perform so that the completion time of a ship operation is minimized. Fig. 1
provides a drawing of QCs working on a ship.
Inbound containers with the same loading port, of the same size, and transported by the same ship are
said to be included in the same container group. Likewise, outbound containers with the same destination
port, of the same size, and to be loaded onto the same ship are said to be in the same container group. For
sake of efficiency in the discharging and loading operations, in the stowage plan for a ship, a collection of
slots that are located adjacent to each other in the ship are usually allocated to containers of the same
group. In the stowage plan for loading, a cluster is defined to be a collection of adjacent slots into which
containers of the same group are planned to be loaded. In the stowage plan for discharging, a cluster
implies a collection of adjacent slots in which inbound containers of the same group are stacked.
Fig. 2 illustrates a stowage plan for a ship. This plan consists of four cross-sectional views, each cor-
responding to a ship-bay and labeled with an odd number from 1 to 7. Each small square represents a slot.
Shaded squares correspond to slots that containers must be discharged from or loaded onto in this con-
tainer terminals. The shaded pattern in each slot represents a specific group of containers to be loaded into
or picked up from the corresponding slots. Fig. 2 shows that four groups of containers should be dis-
charged from five clusters of slots and then four groups of containers should be loaded into five clusters of
slots.
754 K.H. Kim, Y.-M. Park / European Journal of Operational Research 156 (2004) 752–768

QC1

QC2

QC3

Fig. 1. A drawing of QCs working on a vessel.

7 Cluster 5 3 Ship-bay 1
number number
2
3
Load
Plan
5 1
4

10 7 5 3 1
7

Discharge
Plan
9 8 6

Fig. 2. A partial example of a stowage plan.

This paper defines a ‘‘task’’ as a discharging or loading operation for a cluster. This paper assumes that
once a QC starts to load (or discharge) containers into (from) a cluster of slots, it continues to do so until all
the slots in the cluster become filled (empty). Therefore, this paper considers handling work for a cluster to
be a task.
Similar to the m-parallel machine scheduling problem (e.g., Guinet, 1993; Lee and Pinedo, 1997), the QC
scheduling problem in this study addresses scheduling multiple QCs to perform discharging and loading
operations in order to minimize the makespan of the operation. However, the QC scheduling problem has
several unique characteristics that are different from those of a typical m-parallel machine problem. For
 K.H. Kim, Y.-M. Park / European Journal of Operational Research 156 (2004) 752–768 755

Quay Crane Schedule

QC 1 (operation time: 09:00 ~ 12:00) QC 2 (operation time: 09:00 ~ 12:00)
Operation Cluster Location Type of Number of Start Finish Operation Cluster Location Type Number of Start Finish
sequence number of task task containers time time sequence number of task of task containers time time
1 6 1 Hold* D** 47 09:00 09:47 1 7 3 Deck D 39 09:00 09:39
2 1 1 Hold L** 42 09:47 10:29 2 9 5 Hold D 46 09:41 10:26

3 8 3 Hold D 32 10:31 11:03 3 5 5 Hold L 23 10:26 10:49

4 4 3 Hold L 8 11:03 11:11 4 10 7 Deck D 24 10:51 11:14

5 3 3 Deck L 8 11:11 11:19

6 2 3 Deck L 16 11:19 11:35

* The hold of ship-bay 1.

** D (discharging), L (loading).

Fig. 3. An example of a QC schedule.

example, when discharging and loading operations must be performed at the same ship-bay, the dis-
charging operation must precede the loading operation. When a discharging operation is performed in a
ship-bay, tasks on a deck must be performed before tasks in the hold of the same ship-bay are performed.
Also, the loading operation in a hold must precede the loading operation on the deck of the same ship-bay.
Thus, there are precedence relationships among clusters, relationships that must be observed during a ship
operation. Also, it should be noted that QCs travel on the same track. Thus, certain pairs of tasks cannot be
performed simultaneously when the locations of the two clusters corresponding to the tasks are too close to
each other, because two adjacent QCs must be apart from each other by at least one ship-bay so that they
can simultaneously perform their tasks without interference. Also, if containers for any two tasks must be
picked up at or delivered to the same location in a yard, the two tasks may not be performed simulta-
neously, because doing so will cause interference among yard cranes that transfer containers corresponding
to the two tasks.
Fig. 3 illustrates a QC schedule that shows the number of containers in each cluster, the sequence of
tasks to be performed, and the time schedule for performing the tasks.

3. A mathematical formulation

This section proposes a mathematical formulation for the QC scheduling problem. The constraints in the
scheduling operations of QCs are shown below:

1. Each QC can operate after its earliest available time.

2. QCs are on the same track and thus cannot cross each other.
3. Some tasks must be performed before others.
4. There are some tasks that cannot be performed simultaneously.

The following notations are used for a mathematical formulation.

Indices:
i, j Tasks to be performed. Tasks are ordered in an increasing order of their relative locations in the
direction of increasing ship-bay numbers.
k QCs where k ¼ 1; . . . ; K. QCs are also ordered in an increasing order of their relative locations in
the direction of increasing ship-bay numbers.
756 K.H. Kim, Y.-M. Park / European Journal of Operational Research 156 (2004) 752–768

Problem data:
pi The time required to perform task i.
rk The earliest available time of QC k.
li The location of task i (expressed by the ship-bay number).
l0k The starting position of QC k expressed by a ship-bay number.
lTk The final position of QC k expressed by a ship-bay number. This final position may be specified if
a job for the next ship is already assigned to QC k.
k k
tij The travel time of a QC from location (li ) of task i to location (lj ) of task j. t0j and tjT respectively
represent the travel time from the initial position (l0k ) of QC k to location (lj ) of task j, and from
location (lj ) of task j to the final destination (lTk ) of QC k.
M A sufficiently large constant.
a1 The weight for the makespan (the maximum completion time).
a2 The weight for the total completion time.

Sets of indices:
X The set of all tasks.
W The set of pairs of tasks that cannot be performed simultaneously. When tasks i and j cannot be
performed simultaneously, ði; jÞ 2 W.
U The set of ordered pairs of tasks between which there is a precedence relationship. When task
i must precede task j, ði; jÞ 2 U.

Decision variables:
Xijk 1, if QC k performs task j immediately after performing task i; 0, otherwise. Tasks 0 and T will be
considered to be the initial and final states of each QC, respectively. Thus, when task j is the first
task of QC k, X0jk ¼ 1. Also, when task j is the last task of QC k, XjTk ¼ 1.
Yk The completion time of QC k.
Di The completion time of task i.
Zij 1, if task j starts later than the completion time of task i; 0, otherwise.
W Time at which all tasks are completed.

The QC scheduling problem can be formulated as follows:

X
K
Minimize a1 W þ a2 Yk ; ð1Þ
k¼1

subject to
Yk 6 W 8k ¼ 1; . . . ; K; ð2Þ
X
X0jk ¼ 1 8k ¼ 1; . . . ; K; ð3Þ
j2X
X
XiTk ¼ 1 8k ¼ 1; . . . ; K; ð4Þ
i2X
XX
Xijk ¼ 1 8j 2 X; ð5Þ
k i2X
X X
Xijk  Xjik ¼ 0 8i 2 X; 8k ¼ 1; . . . ; K; ð6Þ
j j

Di þ tij þ pj  Dj 6 Mð1  Xijk Þ 8i; j 2 X; 8k ¼ 1; . . . ; K; ð7Þ

 K.H. Kim, Y.-M. Park / European Journal of Operational Research 156 (2004) 752–768 757

Di þ pj 6 Dj 8ði; jÞ 2 U; ð8Þ
Di  Dj þ pj 6 Mð1  Zij Þ 8i; j 2 X; ð9Þ
Zij þ Zji ¼ 1 8ði; jÞ 2 W; ð10Þ
X k X k X
X
Xujv  Xuiv 6 MðZij þ Zji Þ 8i; j 2 X; li < lj ; 8k ¼ 1; . . . ; K; ð11Þ
v¼1 u2X v¼1 u2X
k
Dj þ tjT  Yk 6 Mð1  XjTk Þ 8j 2 X; 8k ¼ 1; . . . ; K; ð12Þ
k
rk  D j þ t0j þ pj 6 Mð1  X0jk Þ 8j 2 X; 8k ¼ 1; . . . ; K; ð13Þ
Xijk ; Zij ¼ 0 or 1 8i; j 2 X; 8k ¼ 1; . . . ; K; ð14Þ
Yk ; D i P 0 8i 2 X; 8k ¼ 1; . . . ; K: ð15Þ
In the objective function (1), it is assumed that a1  a2 , because the minimization of the makespan is
considered to be more important than the minimization of the total completion time. This is a valid as-
sumption because a ship can depart only after every QC assigned to it completes all the assigned tasks, and
the earliest departure is the primary objective of QC scheduling. However, among possible solutions having
the minimum makespan, the schedule with the shortest total completion time must be selected so that more
QCs may become available to other ships as soon as possible. Constraint (2) evaluates the makespan.
Constraints (3) and (4) respectively select the first and last tasks for each QC. Constraint (5) ensures that
every task must be completed by exactly one QC. Constraint (6) is a flow balance constraint, guaranteeing
that tasks are performed in well-defined sequences. Constraint (7) simultaneously determines the completion
time for each task and eliminates sub-tours. When required, constraint (8) denotes that task i should be
completed before task j. Constraint (9) defines Zij such that Zij ¼ 1 when the operation for task j starts after
the operation for task i is completed; 0, otherwise. Constraint (10) guarantees that tasks i and j cannot be
performed simultaneously when ði; jÞ 2 W. By constraint (11), interference among QCs can be avoided.
Suppose that tasks i and j are performed simultaneously and li < lj . This means that Zij þ Zji ¼ 0. Note that
both QCs and tasks are ordered in an increasing order of their relative locations in the direction of increasing
ship-bay number. Suppose that, for k1 < k2 , QC k1 performs tasks j and Pk1 QCPk2 performs Pk1task
Pi. Then, in-
v v
terference between QCs k1 and k2 results. However, in such a case, v¼1 u2X Xuj  v¼1 u2X Xui ¼ 1,
which cannot be allowed because of constraint (11), and Zij þ Zji ¼ 0. The completion time of each QC is
defined by constraint (12). Constraint (13) restricts the earliest starting time of operations by each QC.
The formulation of constraints (1)–(15) is a mixed-integer linear program. A computational experiment
showed that the computational time is excessive for practical use. For example, a problem with two QCs
and six tasks required 480 minutes to solve completely, using extended LINDO/PC (Release 6.1), which is
a specialized software for integer and linear programs.

4. A branch and bound method

In the B & B method of this study, a solution is represented as a sequence of all tasks. Once a sequence of
all tasks is given, a QC schedule can be constructed by assigning tasks to QCs in such a way that the first
unassigned task in the sequence is assigned to the QC with the earliest completion time for already assigned
tasks. This approach is called a ‘‘list scheduling.’’ However, when evaluating the completion time of a QC,
a delay of a QC resulting from interference with another QC must be considered. In list scheduling, no
insertion of idle time is assumed. When compared to the formulation in Section 3, the solution space of this
section is reduced by the above assumptions.
The B & B method essentially explores a solution tree starting from the root node at level 0 through
nodes in the higher levels. The number of levels in the solution tree for the QC scheduling problem is the
758 K.H. Kim, Y.-M. Park / European Journal of Operational Research 156 (2004) 752–768

same as the number of all tasks. At each level, a task is selected during the enumeration process. Therefore,
a node at level i in the search tree corresponds to a partial solution that consists of the 1st through the ith
tasks to be performed.
The following notations were used to describe the solution procedure.

Wk The set of tasks assigned to QC k.

Sk The sequence of tasks assigned to QC k.
ck The completion time of QC k at the current node in the search tree.
lck The position of QC k after the last task in Sk of the current node is performed.

At the root node of the search tree, for all k, Wk ¼ /, Sk ¼ ð Þ, ck ¼ 0, and lck ¼ l0k .
At each node, a solution is represented as fSk g1 6 k 6 K . The overall procedure is shown in Fig. 4. A
detailed explanation for each step is provided below.

Select the next branching node until the first feasible solution is found
Until the first feasible solution is found, a depth-first search is performed. That is, the next node for
branching is selected among the newly sprouted nodes. The node with the minimum lower bound among
the newly sprouted nodes is selected as the next branching node. The depth-first search is continued until
the first feasible solution is found. The objective value of the first feasible solution becomes the initial value
of the upper bound of the optimal objective value.

Select the next branching node after the first feasible solution is found
Once the first feasible solution is found by the depth-first search, all the nodes with partial solutions will
be candidates for the next branching. The selection criteria for the next branching node is still the minimum
lower bound.

10. Select the schedule with

1. Initialize the least total completion
time among schedules with
the shortest makespan
No
2. Select the next 9. No more
branching node Yes
unexplored nodes?

3. Make the feasible

children nodes from the
selected branching node

4. Delete dominated
nodes 8. Update the upper
bound

5. Calculate the lower

bounds for children
nodes Yes
No

6. Fathom nodes by
using the lower bounds 7. Last level?

Fig. 4. The overall procedure of the B & B method.

 K.H. Kim, Y.-M. Park / European Journal of Operational Research 156 (2004) 752–768 759

Make feasible children nodes from the branching node

First, among Sj , j ¼ 1; . . . ; k, the sequence, Sk , with the minimum ck is selected. Then, each remaining
task will be appended to the sequence to make a child node. In the process, remaining tasks that violate the
precedence constraints between tasks (constraint (8)) or the constraints regarding the interference between
QCs (constraints (10) and (11)) are deleted.

Delete dominated nodes

Suppose that more than one node has the same elements of [Wk and the value of lck for all k. Then, node
a is said to dominate node b if node a has a value of ck less than or equal to that of node b for all k, and, for
at least one k, node a has a value of ck strictly lower than that of node b. All nodes dominated by another
node are deleted.

Calculate the lower bound of children nodes

For each node, a lower bound is calculated and compared with the current upper bound of the optimal
objective value. If the lower bound is greater than the current upper bound, the corresponding sub-tree
connected to the node will no longer be explored.
When the locations of QCs and remaining tasks are given, an interval of ðlck ; lckþ1 Þ is divided into several
partitions by the positions of the remaining tasks, as shown in Fig. 5. Let the set of remaining tasks be
U ð¼ X  [k2C Wk Þ and the length of the ith partition be denoted as uki , which is expressed in units of travel
time of QCs. Then, for a node in the search tree, a lower bound on the optimal objective value can be
expressed as follows:
Lower bound 1 ¼ maxfmax ck ; bm g; ð16Þ
k

where
( )
1 X X
bm ¼ ck þ pi þ QCtravel ; ð17Þ
K k2C i2U

QCtravel ¼ tl þ tb þ tr ; ð18Þ

tl ¼ max flc1  lj ; 0g; ð19Þ

j2U

tr ¼ max flj  lcK ; 0g; ð20Þ

j2U

X
K 1
tb ¼ uk ð21Þ
k¼1

tl u11 u12 u13 u21 u22 tr

Fig. 5. Illustration of notations for defining the lower bound of the optimal objective value.
760 K.H. Kim, Y.-M. Park / European Journal of Operational Research 156 (2004) 752–768

and
X
uk ¼ uki  max fuki g: ð22Þ
i
i

The following explains why expression (16) can be a lower bound on the optimal objective value. The
value of tl is the lower bound on the travel time of a QC required to perform the left-most task, while tr is
that required to perform the right-most task. Let an interval ðlck ; lckþ1 Þ be divided into m partitions according
to the different positions of remaining tasks. Then, to complete the tasks between consecutive QCs, at least
(m  1) partitions must be covered by QC movements. Because uk is the minimum sum of lengths of (m  1)
partitions, uk becomes the lower bound on the total travel time of QCs required for performing tasks in the
interval ðlck ; lckþ1 Þ. Thus, Eq. (21) is the lower bound on the travel time of QCs for completing all the re-
maining tasks between QC k and (k þ 1). Therefore, P l þ tr þ tb ) is the lower bound on the travel time to
(tP
complete all the remaining tasks. Thus, f k2C ck þ i2U pi þ tl þ tr þ tb g is less than or equal to the
minimum total completion time of QCs. Therefore, the optimal makespan must be greater than or equal to
the total completion time, divided by the number of QCs. Thus, expression (16) is a lower bound on the
optimal objective value.
Lower bound 2 is the same as lower bound 1 except that QCtravel is evaluated as follows. While a lower
bound on the minimum travel time of QCs is used as QCtravel for lower bound 1, for lower bound 2, the exact
minimum travel time of QCs is used as QCtravel . This problem of finding the exact minimum travel time is
equivalent to the multiple traveling salesmen problem (MTSP). Because QCtravel for lower bound 2 is the
exact minimum travel time, lower bound 2 is tighter than lower bound 1. To find the optimal solution of
MTSP, tasks must be optimally allocated to one of the QCs, and the optimal travel route for each QC for
visiting the locations of all the allocated tasks must be determined.
By considering that all the tasks and QCs are located on a straight line, the optimal allocation of tasks to
QCs has the form of solutions that can be obtained as follows. Partition each interval ðlck ; lckþ1 Þ into left and
right sub-intervals. Then, allocate tasks on the left sub-interval to QC k, and tasks on the right sub-interval
to QC (k þ 1). Thus, to find the optimal allocation, it is sufficient to enumerate all the possible partition
points––which results in different allocations of tasks to QCs––between adjacent QCs. After all tasks are
allocated to QCs, minimize the travel time as follows:
X
K
QCtravel ¼ tk ; ð23Þ
k¼1

where
tk ¼ maxfðlck  lkl Þ; ðlkr  lck Þg þ 2 minfðlck  lkl Þ; ðlkr  lck Þg: ð24Þ

In expression (24), lkl is the location of the left-most task among tasks in the left-hand side of the lo-
cation of QC k, while lkr is the location of the right-most task among tasks in the right-hand side of
the location of QC k. Suppose that lck  lkl 6 lkr  lck . Then, QC k can minimize its travel time by moving
to the location of the left-most task first and then moving to the location of the right-most task. In such a
case, the travel time of QC k will be ðlkr  lck Þ þ 2ðlck  lkl Þ. When lck  lkl > lkr  lck , the travel time of QC k
will be ðlck  lkl Þ þ 2ðlkr  lck Þ. Note that lkl and lkr become lck when there is no allocated task in the left-hand
side or the right-hand side of QC k, respectively.

Stop the search process

When there remains no more nodes with a partial solution, stop the process. Select the schedule with the
smallest makespan. If there are multiple schedules with the smallest makespan, then select among them the
schedule with the least total completion time.
 K.H. Kim, Y.-M. Park / European Journal of Operational Research 156 (2004) 752–768 761

The B & B method provides the optimal solution within a short computational time when the number of
tasks is relatively small. However, the computational time increases rapidly when the problem size in-
creases, which makes the B & B method difficult to use in practice. Therefore, GRASP, which is a heuristic
procedure, is suggested, as discussed in Section 5.

5. A quay cranes scheduling procedure using greedy randomized adaptive search procedure

GRASP was developed in the late 1980s to solve combinatorial problems with high complexity as the set
covering problem (Feo and Resende, 1995). Since then, it has been applied to a variety of combinatorial
problems such as scheduling, graph-related problems, quadratic assignment problems, crew scheduling,
machine and tool assignments, and location problems. The overall procedure of GRASP is summarized in
Fig. 6.
GRASP consists of two phases: the solution construction phase and the solution improvement phase. In
the solution construction phase, an element of a solution is added iteratively to an incomplete solution, and
the value of the element is selected by using a random number and a greedy function. In the solution
improvement phase, the constructed solution is locally improved until no more improvement is possible.
Iterations are repeated a prespecified number of times. Then, the best solution is selected among all the
obtained solutions. GRASP can be applied to the QC scheduling problem, as follows:

Solution construction phase (Phase 1)

Step 1: Among all QCs, the QC with the minimum ck is selected. A tie is broken randomly. Let QC i be
selected in this step.
Step 2: A set of feasible tasks for the next operation of QC i is constructed. Let the set be denoted as F .
Tasks that violate various constraints (the precedence constraint between tasks or interference
among QCs) are excluded from F . Also, all the tasks j 2 F for which the following condition
holds are also excluded from F :
1=jlci  lj j < r max f1=jlci  ll jg; ð25Þ
l2F

where r is a constant between 0 and 1.

Step 3: Associate task j 2 F with the following probability:

Fig. 6. A pseudo-code for GRASP.

762 K.H. Kim, Y.-M. Park / European Journal of Operational Research 156 (2004) 752–768

X
ð1=jlci  lj jÞ= ð1=jlci  ll jÞ: ð26Þ
l2F

Step 4: Randomly select a task j 2 F based on the probability distribution defined by (26).
Step 5: Steps 1–4 are repeated until all the tasks are assigned.

According to Eqs. (25) and (26), tasks that are located far from the location of a QC will be excluded
from candidates for the next assignment to the QC (in case of (25)), or the tasks will have a low probability
of being selected for the next assignment to the QC (in case of (26)).

Solution improvement phase (Phase 2)

Step 0: i ¼ 0.
Step 1: i ¼ i þ 1. If i > K, stop. Otherwise, select QC i. Go to step 2.
Step 2: Select a pair of tasks that are assigned to QC i and for which the highest improvement in
maxk fck g can be obtained by performing a pairwise interchange (the 2-opt method). If there
is more than one pair of tasks with the same improvement in maxk fck g, select the task with
the minimum total completion time. If the improvement is positive, perform the interchange.
Repeat this step until no further improvement is possible. If no further improvement is possible,
go to step 1.

Note that, even when only two tasks for the same QC is interchanged, the completion time of all the QCs
must be evaluated again because the interference among QCs must be checked again. Even though more
complicated improvement procedures (for example, 3-opt or interchange of tasks among different QCs)
may be applied, the long computational time caused by the need to consider the interference among QCs
make them feasible.

The stopping criteria

The solution construction procedure and the solution improvement procedure are performed repeatedly
until the maximum number of iterations is reached or no better solution is found during a specified number
of iterations (SNIs). When the stopping criteria is satisfied, the schedule with the minimum total completion
time, among schedules with the shortest makespan, is selected as the final solution.
A numerical example of the QC scheduling procedure using GRASP is provided in the following. The
input data from Table 1 was used.

Phase 1

(Step 1) QC 1 is selected ðc1 ¼ 0; c2 ¼ 0Þ.

(Step 2) X ¼ f1; 2; 3; 4; 5g, F ¼ f1; 2; 3g
1=jlc1  l1 j ¼ 1=1 > 0:4 maxf1=1; 1=2; 1=3g ¼ 0:4
1=jlc1  l2 j ¼ 1=2 > 0:4 maxf1=1; 1=2; 1=3g ¼ 0:4

Table 1
Input data for the example
Task (i) 1 2 3 4 5
pi 8 10 20 5 30
li 1 2 3 5 6
W ¼ fð1; 2Þ; ð2; 3Þ; ð3; 4Þ; ð4; 5Þg, U ¼ f g, K ¼ 2, lc1 ¼ 0, lc2 ¼ 4, r ¼ 0:4
 K.H. Kim, Y.-M. Park / European Journal of Operational Research 156 (2004) 752–768 763

1=jlc1  l3 j ¼ 1=3 < 0:4 maxf1=1; 1=2; 1=3g ¼ 0:4

F ¼ f1; 2g
(Step 3) According to Eq. (26), the probability of two tasks become as follows:
ðð1=1Þ=ð1=1 þ 1=2Þ; ð1=2Þ=ð1=1 þ 1=2ÞÞ ¼ ð2=3; 1=3Þ
(Step 4) Let the generated random number be 0.52. Job 1 is selected.
c1 ¼ 1 þ 8 ¼ 9; c2 ¼ 0; lc1 ¼ 1; lc2 ¼ 4
(Step 1) QC 2 is selected ðc1 ¼ 9; c2 ¼ 0Þ.
(Step 2) X ¼ f2; 3; 4; 5g, F ¼ f3; 4; 5g
1=jlc2  l3 j ¼ 1=1 > 0:4 maxf1=1; 1=1; 1=2g ¼ 0:4
1=jlc2  l4 j ¼ 1=1 > 0:4 maxf1=1; 1=1; 1=2g ¼ 0:4
1=jlc2  l5 j ¼ 1=2 > 0:4 maxf1=1; 1=1; 1=2g ¼ 0:4
F ¼ f3; 4; 5g
(Step 3) According to Eq. (26), the probability of tasks being selected become as follows:
ðð1=1Þ=ð1=1 þ 1=1 þ 1=2Þ; ð1=1Þ=ð1=1 þ 1=1 þ 1=2Þ; ð1=2Þ=ð1=1 þ 1=1 þ 1=2ÞÞ ¼ ð2=5; 2=5; 1=5Þ
(Step 4) Let the generated random number be 0.71. Job 4 is selected.
c1 ¼ 9; c2 ¼ 1 þ 5 ¼ 6; lc1 ¼ 1; lc2 ¼ 5
(Step 1) QC 2 is selected ðc1 ¼ 9; c2 ¼ 6Þ.
...
...
(Step 5) All the tasks are assigned to QCs. The sequences of tasks for QCs 1 and 2 are S1 ¼ ð1; 3; 2Þ and
S2 ¼ ð4; 5Þ, respectively.

Phase 2

(Step 1) QC 1 is selected.
(Step 2) The neighboring sequence with the largest improvement is S1 ¼ ð1; 2; 3Þ. No more improvement is
possible. Go to step 1.
(Step 1) QC 2 is selected.
(Step 2) The neighboring sequence with the largest improvement is S2 ¼ ð5; 4Þ. No more improvement is
possible. Stop phase 2. The first iteration is ended.

6. A numerical experiment

Three different experiments were performed. The first experiment compared two lower bounds for the B
& B method. Twenty-two problems (p13–p34 in Table 2) were solved by using lower bounds 1 and 2. Fig. 7
shows the ratio of the performance (the number of nodes enumerated and the computational time) of the B
& B method using lower bound 2 to that using lower bound 1. Although the B & B method using lower bound
2 outperformed that using lower bound 1 based on the number of nodes enumerated, their computational
times were similar. This is because lower bound 2 requires more time to evaluate.
The second experiment searched for the best set of parameters, (SNI, r) for GRASP, and the third
experiment was to compared the performance of the B & B method with lower bound 1 with that of
GRASP. For the second experiment, 12 problems (p1–p12), which have 2–3 QCs and 10–20 tasks, were
generated. It was assumed that the number of ship-bays in the corresponding vessel was the same as the
number of tasks. The initial locations of QCs were assumed to be equally spaced on the vessel. The op-
eration time for a task was randomly generated from a uniform distribution of U ð3; 180Þ. Also, attributes
of tasks (hold task or deck task, loading or discharging) were chosen randomly. The precedence rela-
tionships among tasks in the same ship-bay––discharging tasks on deck, discharging tasks in a hold,
loading tasks in a hold, and loading tasks on deck––were considered.
764 K.H. Kim, Y.-M. Park / European Journal of Operational Research 156 (2004) 752–768

Table 2
The results of the computational experiment for problems p13–p49 (r ¼ 0:7, SNI ¼ 50)
Test problems B&B GRASP B & B/GRASP
Problem The number Number (1) Make- (2) Computation (3) Make- (4) Computation (3)/(1) (4)/(2)
no. of QCs of tasks span time (seconds) span time (seconds)
13 2 10 453 20 453 28 1.00 1.40
14 2 10 546 17 546 17 1.00 1.00
15 2 10 513 13 516 11 1.01 0.85
16 2 10 321 64 321 15 1.00 0.23
17 2 10 456 28 456 23 1.00 0.82
18 2 10 375 35 375 31 1.00 0.89
19 2 10 552 44 552 22 1.00 0.50
20 2 10 480 17 480 22 1.00 1.29
21 2 10 465 1 465 16 1.00 16.00
22 2 10 720 22 720 26 1.00 1.18
23 2 15 576 372 591 65 1.03 0.17
24 2 15 669 368 675 94 1.01 0.26
25 2 15 738 2279 741 88 1.00 0.04
26 2 15 639 78 651 126 1.02 1.62
27 2 15 657 187 687 52 1.05 0.28
28 2 15 537 4010 549 83 1.02 0.02
29 2 15 807 651 819 94 1.01 0.14
30 2 15 891 12 906 97 1.02 8.08
31 2 15 570 2468 570 57 1.00 0.02
32 2 15 591 91 597 121 1.01 1.33
33 3 20 603 8973 666 110 1.10 0.01
34 3 20 717 8877 762 118 1.06 0.01
35 3 20 678–690a 41014 699 127 0.00
36 3 20 678–720 46355 708 238 0.01
37 3 20 510–516 43961 540 279 0.01
38 3 20 609–633 38749 660 160 0.00
39 3 20 513–552 47125 579 234 0.00
40 3 20 558–576 34473 597 148 0.00
41 3 20 585–654 33893 642 269 0.01
42 3 20 561–588 35264 666 214 0.01
43 3 25 864–951 48938 942 291 0.01
44 3 25 813–879 39649 858 619 0.02
45 3 25 825–861 57568 873 399 0.01
46 3 25 687–708 24506 735 667 0.03
47 3 25 789–912 43687 807 437 0.01
48 3 25 627–669 55966 669 307 0.01
49 3 25 888–915 69771 972 457 0.01
a
An optimal solution could not be obtained because the number of nodes searched at a level exceeded 175,000, the preset limit in the
program.

The GRASP algorithm was programmed by using Microsoft C 6.0 language on a Pentium II-466 MHz
PC with 64 MB RAM. The algorithm was stopped when the number of iterations reached 500 or when no
better solution was found during SNI iterations. In the experiment, the SNI was set to 10, 25, 50, 100, or
200. The value of r in Eq. (25) was set to 0.1, 0.3, 0.5, 0.7, or 0.9. The experiment was conducted under 25
different combinations of parameters from five values of SNI and five values of r.
Among 12 test problems, only nine (p1–p9) could be solved by the B & B method. The nine problems
were solved under 25 different experimental conditions. Fig. 8 shows that there was little change in the
makespans when r was less than 0.5. The makespan was lowest when r was between 0.5 and 0.7. However,
 K.H. Kim, Y.-M. Park / European Journal of Operational Research 156 (2004) 752–768 765

CPU Time # of nodes

1.4
1.2
1
0.8
0.6
0.4
0.2

0
13 15 17 19 21 23 25 27 29 31 33

Fig. 7. A comparison of the number of enumerated nodes and the computational time for two lower bounds.

1.3 p1
p2
p3
1.2 p4
p5
p6
1.1 p7
p8
p9
1

0.9
0.1 0.3 0.5 0.7 0.9 r
Fig. 8. Normalized makespans for various values of r (SNI ¼ 100).

the makespan increased dramatically when the value of r equaled 0.9. Generally, a better solution was
found with a larger value of SNI. We found that, for a low value of r, the improved amount in the objective
value during the improvement phase increased as the SNI increased. In comparison, when the value of r
was high, the amount of improvement in the objective value was less sensitive to the change in the SNI. The
algorithm became similar to a random search when the value of r was very small. Through a preliminary
experiment for the QC scheduling problem, it was concluded that the best value of r is 0.7 (Fig. 8). Fig. 9
shows that the solution was improved as the value of the SNI increased. However, even at SNI ¼ 10, the
objective value resulting from GRASP did not exceed the optimal objective value by more than 10%. Also,
Fig. 9 shows that the objective values found by GRASP became less than 103% of the optimal objective
values after the SNI exceeded 100. Fig. 10 shows that the computational time was proportional to the value
of the SNI.
766 K.H. Kim, Y.-M. Park / European Journal of Operational Research 156 (2004) 752–768

1.1 p1
p2
p3
1.05 p4
p5
p6
p7
1
p8
p9

0.95 SNI
10 25 50 100 200 B&B
Fig. 9. Makespan for various values of SNI (r ¼ 0:7).

25

20

15

10

0 SNI
0 50 100 150 200

Fig. 10. The average CPU time for various values of SNI (r ¼ 0:7, p1–p9). CPU time at SNI ¼ 10 is normalized to be one.

The third experiment set SNI ¼ 50 and r ¼ 0:7. Ninety problems (p13–p102) were randomly generated as
described previously. The number of QCs and tasks were assumed to be 2–6 and 10–50, respectively.
Table 2 compares makespans and computational times of p13–p49 between the B & B method and
GRASP. For p13–p34, the B & B method obtained the optimal solution. However, for p35–p49, the B & B
method could not obtain the optimal solution within a prespecified time. Thus, the search procedure of the
B & B method was stopped when the number of nodes at a search level exceeded 175,000. For p35–p49 (the
maximum lower bound, the minimum upper bound) obtained thus far were listed in the fourth column, and
the computational times consumed before the stop of the search were listed in the fifth column. When the
problem size is small, i.e., for p13–p22, GRASP also obtained optimal solutions. For problems 23–34, the
objective values from GRASP were higher than those from the B & B method by less than 10%. However,
the computational times of GRASP became less than 3% of those of the B & B method when the number of
QCs and the number of tasks exceeded 3 and 20, respectively.
Based on the results of 90 problems (p13–p102), Fig. 11 shows how the computational time of GRASP
changed as the number of tasks increased. It can be observed that the computational time increased rapidly
as the number of tasks increased. Fig. 12 shows how the computational time changed as the number of QCs
increased. It is interesting to observe that the computational time decreased as the number of QCs in-
creased. Considering that, in practice, more QCs are usually used for larger numbers of tasks, the effect of
the number of tasks on the computational time is offset by the effect of the number of QCs.
 K.H. Kim, Y.-M. Park / European Journal of Operational Research 156 (2004) 752–768 767

4000
3500
3000
2500
2000
1500
1000
500
0
10 15 20 25 30 35 40 45 50

Fig. 11. The average CPU time of GRASP (SNI ¼ 50, r ¼ 0:7, p13–p102).

700

600

500

400 p43
p44
300 p45

200

100

0
2 3 4 5 6

Fig. 12. The CPU time of GRASP for various numbers of QCs (SNI ¼ 50, r ¼ 0:7).

In addition, considering the fact that, in real QC scheduling problems, the number of QCs ranges from 2
to 6 and the number of tasks ranges from 10 to 50, GRASP can be considered as an appropriate meth-
odology for solving practical QC scheduling problems.

7. Conclusion

This paper described a mathematical formulation for the QC scheduling problem, which is an important
problem in the operation of port container terminals. To solve the problem, a B & B method and a lower
bound of the optimal objective value were proposed. In addition, this study proposed the GRASP algo-
rithm, a heuristic search algorithm with a function for escaping from a local minimum, to find near-optimal
solutions to the QC scheduling problem. This study also performed a numerical experiment to evaluate the
performance of the suggested algorithms. The final objective values derived by GRASP did not exceed
those derived by the B & B method by more than 10% when the parameters of GRASP have values within
768 K.H. Kim, Y.-M. Park / European Journal of Operational Research 156 (2004) 752–768

specified ranges. Also, on the average, GRASP reduced the computational times to 3% of those of the B &
B method when the number of QCs and the number of tasks exceeded 3 and 20, respectively. From a
practical perspective, the computation time of GRASP was satisfactory.
It was noted that the ship planning process usually consists of several hierarchical levels of decision
making such as berth planning, QC scheduling, and discharge and load sequencing. The decisions made at a
higher level become constraints for decision making at a lower level. However, there are opportunities to
improve the efficiency of plans by integrating the decision-making process at different levels. The inte-
gration of QC scheduling with berth planning or discharge and load sequencing may be possible with the
help of new information technology, which is a promising topic for future study.

Acknowledgements

The authors would like to thank an anonymous referee for his/her careful review and helpful sugges-
tions. This work was supported by the Korea Science and Engineering Foundation (KOSEF) through the
Center for Intelligent & Integrated Port Management Systems (CIIPMS) at Dong-A University.

References

Brown, G.G., Lawphongpanich, S., Thurman, K.P., 1995. Optimizing ship berthing. Naval Research Logistics 41, 1–15.
Cheung, R.K., Li, C.-L., Lin, W., 2002. Interblock crane deployment in container terminals. Transportation Science 36 (1), 79–93.
Cho, D.W., 1982. Development of a methodology for containership load planning. Unpublished Ph.D. Dissertation. Oregon State
University.
Cojeen, H.P., Dyke, P.V., 1976. The automatic planning and sequencing of containers for containership loading and unloading. In:
Pitkin, Roche, Williams, (Eds.), Ship Operation Automation. North-Holland Publishing Co.
Daganzo, C.F., 1989. The crane scheduling problem. Transportation Research Part B 23B (3), 159–175.
Feo, T.A., Resende, M.G.C., 1995. Greedy randomized adaptive search procedures. Journal of Global Optimization 6, 109–133.
Gifford, L.A., 1981. Containership load planning heuristic for a transtainer-based container port. Unpublished M.Sc. Thesis. Oregon
State University.
Guinet, A., 1993. Scheduling sequence-dependent jobs on identical parallel machines to minimize completion time criteria.
International Journal of Production Research 31 (7), 1579–1594.
Kim, K.H., Kim, K.Y., 1999. An optimal routing algorithm for a transfer crane in port container terminals. Transportation Science 33
(1), 17–33.
Lee, Y.H., Pinedo, M., 1997. Scheduling jobs on parallel machines with sequence-dependent setup times. European Journal of
Operational Research 100 (3), 464–474.
Lim, A., 1998. The berth planning problem. Operations Research Letters 22, 105–110.
Narasimhan, A., Palekar, U.S., 2002. Analysis and algorithms for the transtainer routing problem in container port operations.
Transportation Science 36 (1), 63–78.
Peterkofsky, R.I., Daganzo, C.F., 1990. A branch and bound solution method for the crane scheduling problem. Transportation
Research Part B 24 (3), 159–172.
Zhang, C., Wan, Y.-W., Liu, J., Linn, R.J., 2002. Dynamic crane deployment in container storage yards. Transportation Research Part
B 36, 537–555.